Auto Alpha AdvisoryThree genuinely different products get sold under the single word "automation". The cheapest of them runs on tooling that costs R150 a month. The most expensive needs a person reading its output every day. Work out which one you are buying before you sign anything, because they fail in unrelated ways and the fair price differs by an order of magnitude.
I am a CA(SA) who also writes production code, so I read an automation quote twice: once at the line items, once as the person who would have to build the thing. Here is the part a vendor will not open with. A large share of what gets quoted as an "automation project" is a Make scenario that a competent internal ops person could build in a fortnight. That does not make the quote dishonest. It does mean you are buying judgement rather than software, and the invoice should say which.
The base rate justifies the suspicion. S&P Global's enterprise survey found the proportion of companies abandoning most of their AI initiatives before production has climbed to 42%, from 17% a year earlier, and the average organisation now scraps 46% of its proofs of concept before they reach production.
RPA: a robot clicking your screen, and what it costs when the screen changes
Robotic process automation drives the user interface of software that has no usable API. It types into fields and clicks buttons the way a person would. It works, and it breaks the day a vendor moves a button.
Entry pricing looks harmless. UiPath's Automation Cloud Basic tier starts at R420 a month for a handful of users and two unattended robots. Everything with real capacity is priced by a salesperson, which is your first signal about the total. Ernst & Young, writing in 2019, said it had seen as many as 30-50% of initial RPA projects fail, and the reasons it gave were methodology and process selection rather than the technology itself.
Buy RPA only when there is genuinely no API and no export. Then ask two things before you sign: what happens when the target screen changes, and who pays for that fix. If the answer to the second one is you, at an hourly rate, price the next three years rather than the build.
Integration plumbing: where most automation quotes actually live
Make, Zapier and n8n move data between systems that already expose APIs. This is where the large majority of small-business automation sits, and the tooling is close to free.
Make's Core plan is R150 a month for 10,000 credits. Zapier's Professional plan starts at R500 a month for 750 tasks, billed month to month. n8n's Starter tier is R380 a month for 2,500 executions billed annually, and its self-hosted community edition is free.
So when a quote for "an automation project" carries a five-figure build fee and the deliverable is eleven Make modules, you are paying for scoping and accountability. That can be entirely worth it. It is worth it when the process is ambiguous, when it touches money or POPIA-relevant personal information, or when it has to keep working after the person who built it resigns. It is not worth it when someone in your ops team already understands the process end to end and needs two afternoons and permission. I wrote up the do-it-yourself version of this in The 5 AI Automations Every Business Should Set Up This Week.
LLM-in-the-loop: what you buy alongside the model
An LLM step returns different output for identical input. That is the product, not a defect, and it changes what you have to buy with it: a test set and a human who reads the result.
The tokens are cheap. Anthropic's own worked example prices 10,000 support conversations on its Haiku model at about R620. Almost none of the real cost is inference. It sits in assembling a held-out set of cases with known-correct answers, measuring the model against them, and keeping a person in the path before anything reaches a customer. Skip that and you have bought a system that is confidently wrong at scale, which is worse than the manual process it replaced.
RAND's study of AI project failure puts the headline rate at more than 80%, twice the failure rate of IT projects that do not involve AI. Non-determinism is a large part of why.
Why scoping discipline beats vendor size
Failure clusters around unclear problem definition, not small suppliers. RAND interviewed 65 experienced data scientists and engineers, and the root cause they named most often was the business leadership of the organisation misunderstanding how to set a project on a path to success — not the technology, and not the size of the supplier. Their conclusion is blunt:
Misunderstandings and miscommunications about the intent and purpose of the project cause more AI projects to fail than any other factor.
The S&P data points the same way. Organisations with below-average failure rates weigh more factors when choosing what to automate: 54% consider data availability, against 42% of the organisations with above-average failure rates, and the better performers weigh 4.9 factors on average versus 4.2. Time-to-deliver showed no correlation with success at all. A vendor whose main pitch is speed is selling the one variable the data says does not matter.
Appetite is not the constraint in South Africa. Microsoft's AI Economy Institute put South African AI usage at 23.1% in the first quarter of 2026, up from 21.1% in the second half of 2025, which puts the country 46th of 147 economies measured. The appetite is there. What is usually missing is a clearly defined problem.
The questions that separate a builder from a deck
Ask these in the first meeting. The answers sort vendors faster than a reference call.
- Show me something you built that is running in production right now, and let me watch it run.
- Which of the three types is this, and why is the cheaper type not enough?
- What is the monthly tool bill, and will it be invoiced to my account rather than marked up through yours?
- What breaks this, and what does a fix cost?
- If there is an LLM step: what is the test set, what accuracy does it reach, and who reviews the output before a customer sees it?
- Who owns the scenarios, the credentials and the code when we stop working together?
- What is the smallest version of this that could be live in two weeks?
The last one is the filter. Someone who ships answers it with a scope and a date. Someone selling a roadmap answers it with a discovery phase. That is the cheap version of a test I have written about before, on why an AI consultant should be able to write the code, and it also explains most of why AI projects stall before they ship.
Pick the one process that irritates somebody in your business every single day. Get a fixed-price quote for that alone, with a working demonstration at the end of it, and hold the rest of your budget back until you have seen it run. If a vendor cannot quote that, they cannot build it, and the deck they send instead will cost you a quarter. My implementation and advisory work is shaped around that first shipped thing rather than a strategy document, which is a deliberate choice and not a modest one.